Behavioral Interviews

What kinds of technical bias have you seen?

By Interview Clozo · Published 27 October 2026 · Last verified 26 October 2026 · 11 min read

What kinds of technical bias have you seen? The main sources are biased data, a proxy label, an average that masks a failing group, and misuse; pick one case and say what you changed. Technical bias is a systematic error in data, a model or a system that makes outcomes worse for some groups than others. The natural follow-up is how you found it, and the strong answer is a per-group measurement, not a hunch.

In this article

Before you start
Why interviewers ask this
Technical bias from first principles
The Staff-level answer
The rubric
Three variants of this question
Common mistakes
Follow-up questions
Sources

Before you start

This page belongs to the behavioral interview guide. The question has two readings. This page reads it as bias in artificial intelligence (AI) and machine learning (ML) systems. It can also mean bias in technical hiring, such as interviewer bias. If you are unsure, ask the interviewer which one they mean before you answer; the third variant below applies the same method to the hiring reading.

Why interviewers ask this

A vague answer costs you the point, because the interviewer cannot tell whether you have seen bias or only read about it. Reported by candidates in leadership and behavioral rounds, the question gives you room to show three things: that you know bias comes in more than one kind, that you look for it with a measurement, and that you act when you find it.

A weak answer says "models are biased if the data is biased" and stops. A strong answer names one kind, gives the case, the number that exposed it and the change you made. A Staff-level answer adds who else you had to convince, because bias is rarely fixed by one person's code. The neighbors of this question are in common interview questions and how to answer them.

Technical bias from first principles

Technical bias: Technical bias is a systematic error in the data, the model or the system around it that makes outcomes worse for some groups of people than for others. It is repeatable, not random noise, and it can arise without anyone intending harm.
Subgroup metric: A subgroup metric is a performance number, such as accuracy or error rate, computed separately for one group of examples instead of for all examples together. Comparing subgroup metrics is how a hidden gap becomes visible.

In plain words

Imagine an exam written only from one textbook. Students who studied another book score lower, though they know as much. The exam is not measuring ability. It is measuring overlap with the textbook. A biased model works the same way: it measures resemblance to its training examples. The analogy breaks in one place. A student can complain about the exam, but a model's mistake is silent unless someone measures it.

Google's course says it plainly: machine learning models are not inherently objective, because people choose and prepare the training data. In one precise sentence, technical bias is a repeatable gap in outcomes between groups that traces back to data, labels, measurement or use.

The four steps below follow how you would find and explain a case, from the symptom back to the cause. The list of kinds is not exhaustive; Google's own inventory says so.

Start by splitting the metric by group, then trace the gap through the data, the label and the use.CC BY 4.0 — source: What kinds of technical bias have you seen?

Step-by-step description

Compute the metric for each group.
Count who is missing.
Ask what the label measures.
Check how people use the output.

Split the metric by group before anything else.
Then trace the gap to the data, the label and the use.

Step 1: Split the metric by group to find the bias

Bias is found by computing the same metric separately for each group, because one overall number can look healthy while one group fails. Google warns that aggregate metrics such as precision, recall and accuracy can hide biases against minority groups. Split the metric by group before anything else.

The Gender Shades study tested three commercial gender classification systems. The maximum error rate for lighter-skinned males is 0.8%, and darker-skinned females are the most misclassified group, with error rates of up to 34.7% (Buolamwini and Gebru, 2018). Those two bounds differ by about 43 times.¹

An illustrative example shows how an average hides this. Suppose 95% of cases fall in one group scored at 99% accuracy and 5% in another scored at 65%. Overall accuracy is 97.3%, yet the smaller group's error rate is 35 times the larger group's.¹ You now know the symptom to look for: a per-group gap. The next steps find its cause.

The same systems erred on at most 0.8% of lighter-skinned males and on up to 34.7% of darker-skinned females.CC BY 4.0 — source: What kinds of technical bias have you seen?

Text description of the diagram

The maximum error rate for lighter-skinned males is 0.8%.
Darker-skinned females are the most misclassified group, with error rates of up to 34.7%.

Step 2: Trace the gap to the data

Biased data means the examples do not reflect the people or events the model will meet. Selection bias is an error that occurs when a data set's examples are chosen in a way that does not reflect their real-world distribution. Reporting bias is a mismatch between how often something appears in the data and how often it happens in the world.

Gender Shades shows the first kind: the two face datasets it examined were overwhelmingly lighter-skinned, at 79.6% and 86.2% (Buolamwini and Gebru, 2018). The authors then built a new dataset balanced by gender and skin type to test fairly.

Google's example of the second kind is a sentiment model trained on book reviews. Because people were less likely to review a book they did not feel strongly about, the data holds mostly extreme opinions.

Input: your data, counted by group. Output: who is missing or over-represented. If the groups are balanced and the gap remains, look at the label.

Step 3: Trace the gap to the label

A label can carry bias even when the data covers everyone, because the label records past decisions or a stand-in for the real target. Historical bias is old data reflecting inequities that existed in the world when it was recorded, and it enters through the target when past decisions are used as the thing the model copies. The widely reported Amazon recruiting case shows skewed data and historical outcomes together.

MIT Technology Review, summarizing Reuters, reports that Amazon began building a system in 2014 to rank job seekers and trained it on ten years of resumes, most of which came from men because tech is male-dominated. The tool would reportedly penalize resumes containing the word "women's", and the company scrapped the project after it developed a preference for male candidates in technical roles. Describe it as a lesson about what a model learns, not as proof of how the tool was used.

The proxy label case is Obermeyer et al. (2019): the algorithm used health costs as a proxy for health needs, so equally sick Black patients got lower risk scores. The authors estimate this cut the number of Black patients flagged for extra care by more than half, and changing the target removed the bias. Input: the column you train on. Output: whether it measures what you claim.

Step 4: Trace the gap to the use and the people

A system can be accurate and still harmful in how people use it. Automation bias is the tendency to favor results from automated systems over results from non-automated ones, regardless of error rates. A reviewer who trusts a score without question turns a small model gap into a large decision gap.

The National Institute of Standards and Technology (NIST) says bias is not only a data problem. It describes systemic bias, where institutions disadvantage certain groups, and human bias, such as using a person's neighborhood to fill in missing information. Its image of an iceberg puts the statistical and computational biases above the water and the human and systemic ones below. One of its authors says organizations often default to overly technical solutions.

NIST's abstract adds that biases can lead to harmful impacts regardless of intent. That is why you check even when nobody meant harm. Then trace the gap to the data, the label and the use, and decide who owns each fix.

Knowing it is half the job. Saying it is the other half.

Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.

Start Free Trial →

The Staff-level answer

Use your own case. The text in brackets is a slot for your facts, not a story to copy. Persuading the people who own the data is the communication part; see interpersonal communication. The structure is situation, task, action, result (STAR), covered in the teamwork STAR guide.